SkillMaster: Autonomous Skill Mastery Framework for LLM Agents
A new training framework called SkillMaster enables LLM agents to autonomously create, refine, and select skills during task solving, moving beyond reliance on external teachers or rules. The framework uses trajectory-informed skill review to propose, update, or retain skills based on episode evidence, and evaluates candidate edits via counterfactual utility on probe tasks.
Key facts
- arXiv paper 2605.08693 introduces SkillMaster
- SkillMaster teaches agents to create, refine, and select skills autonomously
- Uses trajectory-informed skill review for skill updates
- Evaluates skill edits via counterfactual utility on probe tasks
- Aims to internalize skills as agent capabilities rather than external resources
Entities
Institutions
- arXiv